Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 26, 2026Singapore

Collusion Is Not A Secrecy Problem: Testing Five Common Assumptions About Monitoring AI Agents

Daniel Jun Wei Foo · Team NTU

Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Collusion Is Not A Secrecy Problem: Testing Five Common Assumptions About Monitoring AI Agents

Open in a new tab (opens in new tab)
Recording (opens in new tab)Code (opens in new tab)
Share

What I did.

I tested what happens when AI agents check each other’s work and one of them is compromised. The game copies a two-key approval chain: a sender privately learns a target option that is wrong for the client. It then tries to recruit two checker agents (GLM-5.3-Flash) with a hidden signal. A shared bonus pays only if both checkers back the target and the reviewer finds nothing. I ran the game twice with the same agents and rewards. In the Permissive study, the signal travelled in pictures, signalling was allowed, and the reviewer’s reports went back to the agents without penalties. That reviewer was GPT-6.1 Sol at Ultra reasoning. In the Guarded study, the signal travelled in a two-page report. The checkers became independent verifiers, paid for correct work and given a trusted source table. The reviewer was Claude Opus 5.5 at medium effort, and it now saw rewards and voting history and could issue fines. I also had fresh monitors re-judge the rounds with more or less information, and had outside readers try to decode the signal.

What I found.

In the Permissive study the agents invented a code and followed it in 23 of 24 votes, even though only 2 of 36 pictures were drawn as briefed. The reviewer mentioned the code in every report yet never called it collusion. Its reports even taught the colluders; one checker wrote, “keep encoding but vary framing.” In the Guarded study the sender still planted codes in 8 of 15 rounds, and outsiders decoded them in 5 of 8. The reviewer flagged all 8 code rounds, but the checkers followed the code in 0 of 30 votes. The one checker that spotted the code declined it, citing its duty as a verifier. Fines made the sender give up for 7 rounds, but the reviewer’s reports leaked the secret code word 3 times. Monitors given voting history caught collusion in 12 of 12 views, against 0 of 11 from pictures alone. The lesson is that collusion is decided by whether the receivers act on a signal, not by how well it is hidden. Oversight should target the checkers’ incentives and the monitors’ information, and keep detailed findings away from the agents.

Reviews

Judging this Sprint?

Review “Collusion Is Not A Secrecy Problem: Testing Five Common Assumptions About Monitoring AI Agents”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

How much would this matter for the field if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. The receiver-side framing is the most original idea here, and the reviewer leaking the decoded tag in 3 rounds is a useful finding. The two studies change roles, rewards, reviewer and prompt at once, so they can't support five general claims. Test one assumption, change one factor at a time, and run more than one group per arm.

Cite this project

@misc{foo2026collusion,
  title = {{Collusion Is Not A Secrecy Problem: Testing Five Common Assumptions About Monitoring AI Agents}},
  author = {Daniel Jun Wei Foo},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/collusion-is-not-a-secrecy-problem-testing-five-common-assumptions-about-monitoring-ai-agents-6s9t}},
  url = {https://apartresearch.com/sprints/projects/collusion-is-not-a-secrecy-problem-testing-five-common-assumptions-about-monitoring-ai-agents-6s9t}
}

More from AI Collusion Research Sprint

All 392 entriesBrowse all projects

A Sprint is running now

AI x Epistemics Research Sprint · Nov 13 - 15, 2026